Will the AI Boom Push Up Mortgage Rates?

Rates here are snapshots from the last week of August 2026: a 6.66% thirty-year fixed from Freddie Mac and a 4.68% weekly-average ten-year Treasury from the Fed’s H.15 release. The mechanics change far more slowly than the levels.

On the August 25, 2026 episode of the Dwarkesh Podcast, SemiAnalysis founder Dylan Patel put AI infrastructure spending at about $11 trillion from 2024 through 2029, with roughly $5 trillion of it borrowed. He then argued that if AI firms are willing to pay 8% instead of the 5% to 6% they pay today, “that makes everyone else in the economy also pay 250 bps more.”¹ For a family shopping for a house, 250 basis points is the difference between a 6.66% mortgage and a 9.16% one. On a $400,000 loan that is about $694 more a month.

The mechanism he describes is real. An investment boom with very high returns raises the demand for capital, and that is the textbook way the real interest rate the economy settles at moves up. The number is a different matter. Nothing in interest-rate theory, in the research on how much borrowing moves bond yields, or in the record of the last big technology build-out supports a mechanical 250 basis point increase for everyone. A sustained AI build-out probably keeps mortgage rates somewhat higher than they would be in a world without it. That is an upside risk to plan around, and it is compatible with mortgage rates falling from today’s level if inflation, Federal Reserve policy, or the mortgage spread move the other way.

This guide is about the financing leg of the argument: how much borrowing, absorbed by whom, at what price. The companion question of whether AI lowers inflation while raising the neutral rate is covered in Will AI Lower Interest Rates?, and the piece of a mortgage rate the Fed does not control is in The Term Premium.

What the podcast actually claims

The episode is worth taking seriously because its numbers come from a firm that tracks data-center construction site by site. Patel’s modeling has capital spending passing $2 trillion a year by 2028 and summing to about $11 trillion over 2024 to 2029, financed by roughly $6 trillion of internal cash flow and $5 trillion of credit.¹ The rate argument then runs in three steps. AI compute earns so much per megawatt that the labs and hyperscalers would keep borrowing even at 8%. Borrowers who will pay 8% soak up the capital that everyone else was borrowing at 5% or 6%. So everyone else pays more too, and the discount rate on every other asset rises with it.

The 250 basis point figure starts as a hypothetical jump in one hyperscaler’s borrowing cost, from the 5% to 6% range to 8%, and is then applied to the whole economy. Patel flags the leap himself, introducing the estimate with “Dude, this is vibing a number.”¹ That is the right label. The host extends the same logic to sovereign defaults in heavily indebted countries and to a collapse in the valuations of ordinary cash-generating businesses. Each extension inherits the original number, so the place to start is whether the number holds.

How borrowing reaches a mortgage rate

A thirty-year mortgage rate is built from a ten-year Treasury yield plus a spread. Nothing in the chain responds to corporate borrowing directly; it has to move the Treasury yield first, and then only part of that move passes through.

• The ten-year Treasury. Its yield is the market’s expected path of short rates plus a term premium. A wave of private borrowing can lift it by raising the real rate the economy needs to balance saving and investment, or by adding to the supply of long-dated bonds the market has to absorb.
• The mortgage spread. In the last week of August 2026 the thirty-year fixed averaged 6.66% and the ten-year averaged 4.68%, a gap of 1.98 percentage points against a post-2008 norm near 1.7.²³ That spread pays for prepayment risk, servicing, guarantee fees, and rate volatility, and it moves on its own.
• The pass-through. A Dallas Fed model fit to twenty years of data finds the mortgage rate moves about 85% as much as the ten-year and less than 20% as much as the fed funds rate, holding volatility and the primary-secondary spread fixed.⁴

Run backwards, the chain says what a 250 basis point mortgage shock would require: about 290 basis points on the ten-year Treasury at the Dallas Fed pass-through, or a spread blowout on top of a smaller Treasury move. Either would be a larger repricing than the ten-year has seen from any single cause outside the 2022 inflation shock. Does the Fed Really Set Interest Rates? walks the whole chain with a worked decomposition.

Six places the argument breaks

1. Cumulative borrowing is not annual borrowing

Bond yields respond to how much new debt the market has to absorb each year relative to the saving available, not to a multi-year total. Five trillion dollars through 2029 is about $1 trillion a year if spread evenly, and by Patel’s own timeline it is back-loaded toward 2028 and 2029. That figure is also gross: it includes debt that refinances maturing debt, borrowing that replaces spending the same companies would have done anyway, and bonds bought by the same pension funds and insurers that would otherwise have bought other corporate bonds. The Minneapolis Fed’s Alisdair McKay puts total US private investment near $5.5 trillion a year, so data-center spending on the podcast’s scale is a fifth of it, which is large.⁵ But he also notes that private investment as a share of GDP has been roughly flat since 2018, because housing and other construction weakened as data centers grew. So far the boom has changed the composition of investment more than its total, and rates respond to the total.

2. An 8% AI bond is a Treasury yield plus a credit spread

A corporate bond yield is the risk-free rate plus compensation for the issuer’s own credit risk and liquidity. A company that adds hundreds of billions of dollars of debt to build depreciating hardware pays more because its own debt load went up, and that shows up as a wider spread over Treasuries rather than a higher Treasury yield. The podcast moves from “AI firms pay 8%” to “everyone pays 250 basis points more” without separating the two. The test is observable: if AI issuers’ spreads widen while the corporate bond index spread and the ten-year barely move, the market is pricing issuer risk, and a homebuyer’s rate is untouched.⁶

3. The quantity and the rate are set together

The argument holds the $5 trillion fixed while rates rise by 250 basis points. In practice the two adjust to each other. Higher rates cancel the projects with ordinary returns, and the $5 trillion includes gas turbines, transmission lines, and buildings that earn utility-like returns, not just GPUs. Higher rates also reward saving, so households and firms supply more of it, and they draw in foreign capital looking for yield. Patel’s premise that AI compute earns enough to justify 8% may be right for the marginal chip. It does not follow that every dollar of the build-out clears at that rate, and the projects that do not clear are the ones that would have done the crowding out.

4. The United States is not a closed economy

A Fed research note from July 2026 finds that roughly 90% of equipment goods for high-technology sectors come from abroad, mostly from East Asia.⁷ A companion note tracking the build-out in public data finds that in quarters when those imports spike, the drag from net exports of computers and parts offsets much of the gross investment.⁸ The dollars spent on imported chips and servers come back as foreign purchases of US assets, including Treasuries and corporate bonds. That is foreign saving financing part of the boom, and it is exactly the flow the crowding-out story assumes away. Domestic construction and power are financed at home, so the offset is partial. The calculator below has a slider for it.

5. Higher discount rates are one row of the stock story

A higher discount rate lowers the present value of every future cash flow, all else equal. All else is rarely equal in a productivity boom. If AI raises growth, the cash flows of ordinary businesses rise too, and mature companies paying out earnings now have shorter duration than growth stocks whose value sits decades out. A railroad’s valuation is more exposed to a recession than to a real-rate move of a fraction of a point. The podcast’s claim that reliable non-AI businesses would “crater” needs the 250 basis point number to be true first.

6. The sovereign-default analogy rests on a Fed policy shock

The comparison offered is the early 1980s, when developing-country defaults followed a spike in dollar interest rates. That spike was the Federal Reserve under Paul Volcker deliberately pushing the fed funds rate near 20% to break inflation, and the debt crisis that followed was the consequence of a monetary regime change.⁹¹⁰ Private borrowing to build data centers is a different kind of event, and if it did overheat the US economy the Fed would tighten against the inflation, not against the borrowing. Heavily indebted countries are exposed to any rise in dollar rates, from any cause, which argues for watching the ten-year whatever moves it. It says nothing about whether AI will move it by 250 basis points.

Putting a number on it

There is a literature on how much a given amount of borrowing moves long yields, and it comes from government deficits, the case where the bond market must absorb every dollar. Thomas Laubach’s Federal Reserve study, which uses projected rather than current deficits to get around reverse causality, finds roughly 25 basis points on the ten-year rate expected five years ahead per percentage point of projected deficit-to-GDP.¹¹ Deficits do not create productive capital and are not financed by the return on the thing they fund, so this is a generous coefficient to apply to private investment. Using it anyway gives an order of magnitude.

Take the podcast’s $5 trillion, spread it evenly over five years, assume every dollar draws on US saving, and price it at the Laubach coefficient. One trillion dollars a year is 3.1% of a $32.5 trillion economy.¹² At 25 basis points per point of GDP, that is about 77 basis points on the ten-year Treasury and, at the Dallas Fed pass-through, about 66 basis points on a thirty-year mortgage: 6.66% becomes 7.32%, and the payment on a $400,000 loan rises from $2,571 to $2,747, about $177 a month. Let half the borrowing be financed abroad or spent on imports and the move is about 33 basis points, or $88 a month. To reach the podcast’s 250 basis points on the same borrowing, the coefficient would have to be about 95 basis points per point of GDP, nearly four times the deficit estimate, or the domestic borrowing would have to run near $3.8 trillion a year, with none of it offset by foreign saving or cancelled projects.

None of these are forecasts. They are the arithmetic of the podcast’s own inputs run through published coefficients, and they land at a fraction of a point. The calculator lets you change every assumption, including the coefficient.

What the 1990s technology boom did to mortgage rates

The closest precedent is the information-technology build-out of the late 1990s. Real IT investment grew 23.8% a year from 1995 to 2000, five times the pace of other equipment, and added over three-quarters of a percentage point to annual GDP growth, before the 2001 collapse in IT spending subtracted 0.4 point.¹³ Fiber and telecom capacity were famously overbuilt. If a private capex boom mechanically raised rates, it should show here.

Annual averages. Sources: Freddie Mac Primary Mortgage Market Survey and the Federal Reserve H.15 ten-year constant-maturity series, both via FRED.²¹⁴

The heaviest investment years, 1998 and 1999, had the lowest mortgage rates of the decade. Rates rose in 2000 after the Fed raised its target from 4.75% to 6.5% between June 1999 and May 2000 to cool an economy that was running hot, then fell hard when the boom ended.¹⁵ Inflation, Fed policy, a global flight to Treasuries during the Asian and Russian crises, and federal budget surpluses all mattered more to the ten-year than the pace of domestic IT spending. The episode supports a transitional rise in rates when the Fed leans against an overheating boom. It does not show capex moving yields on its own.

Longer samples say the same thing. A Cleveland Fed study of 1914 to 2016 finds the correlation between productivity growth and real interest rates is negative over the full sample and indistinguishable from zero after 1948.¹⁶ Hamilton, Harris, Hatzius, and West conclude that the link between the equilibrium real rate and trend growth is “much more tenuous than widely believed.”¹⁷ A San Francisco Fed decomposition of the neutral rate’s decline since the 1990s assigns productivity “a very minor role” and puts the weight on demographics, global factors, and fiscal spending.¹⁸

What the data say so far

The build-out is already large in the national accounts. The Minneapolis Fed reports capital spending by the five largest data-center investors rising from $200 billion in 2024 toward $1 trillion by 2027, and summarizes the effect as AI “moderately heating up today’s economy while we wait for the likely bumpy rollout of productivity gains.” Three measures of aggregate productivity through early 2026 “do not show a budding productivity boom.”⁵ The Fed’s own tracking note finds that AI-related components contributed meaningfully to quarterly GDP growth from 2025 through the first quarter of 2026, that imports offset much of it, and that sectors with more AI exposure show higher labor productivity growth without that showing up in the aggregate yet.⁸

Bond markets have so far moved the other way from the crowding-out story. Andrews and Farboodi examine yields around 30 major model releases from four frontier labs between November 2022 and December 2025 and find that long-term Treasury and TIPS yields fall around releases the forecasting community reads as good news about AI progress, with average declines exceeding ten basis points at some maturities that persist for roughly six weeks.¹⁹ They cannot say why. Lower expected consumption growth, a savings glut as gains accrue to high-saving households, and a safer US fiscal position all fit the data. Model announcements are not capex announcements, and the sample is small, so the finding does not settle anything. It does mean the most liquid market in the world has not been treating better AI as higher rates.

Set against that is the strongest model-based case for higher rates. Lukasz Rachel’s Brookings paper finds that an AI scenario with productivity growth 0.75 point a year faster for a decade, together with higher markups and a lower labor share, raises the safe neutral rate by more than a percentage point.²⁰ Note the units. That scenario assumes roughly ten times the total-factor-productivity gain that Daron Acemoglu’s task-based estimate allows over the same decade, which tops out at 0.66% cumulative.²¹ The gap between those two inputs is the whole debate, and the one-point rise in the neutral rate belongs to the strong-boom end of it.

Timing matters as much as size. The ECB’s Philip Lane draws the distinction between AI raising the level of productivity, which pushes the neutral rate up during adoption and then lets it settle back, and AI permanently raising the growth rate of productivity, which keeps it higher; he also notes that gains flowing to high-saving capital owners dampen the demand effect, leaving the net effect on the neutral rate uncertain.²² A BIS model adds that AI adoption is initially disinflationary when the productivity gain is not anticipated, and inflationary right away when households and firms spend ahead of it.²³

Three scenarios for mortgage rates

• Build-out without a productivity breakthrough. Data centers, turbines, and transmission get built; measured productivity stays ordinary. Construction demand, electricity prices, and the stock-market wealth effect keep the Fed a little tighter than it would otherwise be, and the ten-year carries a somewhat larger term premium. Mortgage rates run tens of basis points above the no-boom path while the spending lasts. This is the base case in the numbers above.
• A large, permanent productivity acceleration. The neutral real rate rises for good. Rachel’s roughly one-point result is the strong version. Mortgage rates would settle higher than the 2010s, though the same boom cools inflation, which takes some of it back through the nominal channel.
• Overbuilding, weak monetization, or a labor shock. Capex is cut the way IT spending was in 2001, precautionary saving rises, and rates fall. The 1990s table is what this looks like.

A 250 basis point economy-wide move belongs in none of these. It needs the borrowing to be several times larger than projected, all of it financed domestically, with no project cancellations and no foreign inflow, and the mortgage spread cooperating. The mechanism deserves a place on the list of upside risks. The number does not.

What to do about it

If you are buying or refinancing

• Qualify and budget at the rate you are quoted today. Treat a future refinance as a bonus, not the plan.
• Stress-test the payment at least one percentage point higher. On a $400,000 loan, 6.66% to 7.66% is about $270 more a month. If that number breaks the budget, the house is too big for the plan whether or not AI moves rates. Are You About to Become House Poor? has a fuller version of that test.
• Keep a down payment you will need within a few years in Treasury bills, a money market fund, or a short-term CD. If rates rise, a long-duration bond fund loses value at the moment you need the cash; stocks can do worse.
• An existing fixed-rate mortgage is insulated. Adjustable-rate loans, home equity lines, and any purchase or refinance you have not locked are exposed.

If you are investing

• Do not sell bonds because of a podcast. Rising yields mark down existing long bonds once and raise the return on everything bought afterward. Match duration to when you need the money; Bonds: Diversifier or Hedge? covers which shocks bonds protect against.
• TIPS protect against unexpected inflation, and their prices still fall when real yields rise. If the AI boom lifts real rates, a TIPS fund takes the same mark-to-market hit as a nominal fund of similar duration.
• A globally diversified index already holds the AI winners at market weight. Overweighting them as a hedge against your own mortgage adds concentration to a household that is already exposed to the same trend through its job.

Key takeaways

• The mechanism is sound; the number is not. A high-return investment boom raises the demand for capital and can lift real rates. The podcast’s 250 basis points is a hypothetical jump in one issuer’s borrowing cost applied to everyone, introduced by its author as “vibing a number.”
• Run through published coefficients, the podcast’s own inputs land at a fraction of a point. $5 trillion over five years, all domestic, at the Laubach deficit elasticity and the Dallas Fed pass-through, is about 66 basis points on a mortgage; with half financed abroad, about 33.
• The 1990s IT boom did not raise rates on its own. The heaviest investment years had the decade’s lowest mortgage rates; rates rose when the Fed tightened and fell when the boom ended.
• Bond markets have moved the other way so far. Long Treasury and TIPS yields fell around good-news AI model releases in 2022 to 2025.
• Plan for an upside risk. Qualify at today’s rate, stress-test a point higher, keep near-term cash short, and do not restructure a portfolio around a rate forecast.

Frequently asked questions

Will the AI boom raise mortgage rates?

Probably somewhat, relative to a world without it, while the build-out is under way. AI investment raises the demand for capital, and the strongest model-based estimate puts the long-run effect on the neutral real rate at about a percentage point under an aggressive productivity scenario. Mortgage rates could still fall from today’s level if inflation, Fed policy, or the mortgage spread move the other way.

Could AI push mortgage rates up by 2.5 percentage points?

Not from the borrowing figures on offer. Five trillion dollars over five years priced at the coefficient estimated for federal deficits is worth roughly two-thirds of a point on a mortgage if every dollar is financed at home, and about a third if half is financed abroad. A 2.5 point move would need several times the borrowing, or a rate sensitivity nearly four times the deficit estimate, with no offsets.

How does corporate borrowing affect mortgage rates?

Only through the ten-year Treasury. Corporate borrowing has to raise the real rate the economy needs to balance saving and investment, or add enough long-dated supply to lift the term premium. About 85% of a ten-year move then passes to the thirty-year mortgage. A company paying more because its own debt load rose shows up as a wider credit spread, which does not touch mortgage rates.

What did the dot-com boom do to mortgage rates?

Thirty-year mortgage rates averaged 7.93% in 1995, fell to 6.94% in 1998 at the peak of the IT build-out, rose to 8.05% in 2000 after the Fed raised its target by 175 basis points, and fell to 6.97% in 2001 when IT investment collapsed. Fed policy, inflation, and global capital flows drove the moves; the pace of capex did not.

Should I wait to buy a house until the AI rate picture is clearer?

No forecast will be clear enough to time a purchase on. Buy when the payment fits your budget at today’s rate with a one-point cushion, and when you expect to stay long enough for the transaction costs to amortize. If the payment only works at a refinanced rate, it does not work.

Related guides

• Will AI Lower Interest Rates?: the inflation versus neutral-rate debate, and why experts split on whether AI complements or replaces labor.
• The Term Premium: what the extra yield on a ten-year pays for, and how deficits and supply feed into it.
• Does the Fed Really Set Interest Rates?: the full chain from the policy rate to a mortgage quote.
• Are We in an AI Bubble?: how to read the investment boom without making a timing bet.
• Rent vs. Buy: the hold-period math that decides a purchase at any rate.

Sources

1. Dwarkesh Patel and Dylan Patel, “Dylan Patel: Anthropic & OpenAI will have most of the world’s compute by 2028,” Dwarkesh Podcast (August 25, 2026); $11 trillion of capex 2024 to 2029, $5 trillion of credit, “a 250 bps increase,” and “this is vibing a number.” dwarkesh.com
2. Freddie Mac, Primary Mortgage Market Survey; 6.66% for the week of August 27, 2026, and annual averages 1995 to 2001 via FRED series MORTGAGE30US. freddiemac.com
3. Board of Governors of the Federal Reserve System, Statistical Release H.15, “Selected Interest Rates” (release of September 1, 2026); ten-year constant maturity, weekly average for the week ending August 28, 2026. federalreserve.gov
4. Matthew McCormick and Srini Ramaswamy, “What drives mortgage rates and their response to monetary policy changes,” Federal Reserve Bank of Dallas (May 7, 2026). dallasfed.org
5. Jeff Horwich, “How is AI influencing interest rates? Investment, productivity, prices, and more,” Federal Reserve Bank of Minneapolis (August 3, 2026), quoting Monetary Advisor Alisdair McKay. minneapolisfed.org
6. ICE BofA US Corporate Index Option-Adjusted Spread, via FRED series BAMLC0A0CM; the series to watch for whether AI issuance is widening corporate spreads rather than moving Treasuries. fred.stlouisfed.org
7. Giuseppe Fiori, Colleen Lipa, and Erik Nuenninghoff, “Technology Shocks, the AI Boom, and the U.S. Current Account,” FEDS Notes, Board of Governors (July 14, 2026); “approximately 90 percent of equipment goods for high-technology sectors originating abroad.” federalreserve.gov
8. Jeffrey S. Allen, Paul E. Soto, and Mason Thieu, “The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact,” FEDS Notes, Board of Governors (July 17, 2026). federalreserve.gov
9. Federal Reserve History, “Volcker’s Announcement of Anti-Inflation Measures, October 1979.” federalreservehistory.org
10. Federal Reserve History, “Latin American Debt Crisis of the 1980s.” federalreservehistory.org
11. Thomas Laubach, “New Evidence on the Interest Rate Effects of Budget Deficits and Debt,” FEDS Working Paper 2003-12; published in the Journal of the European Economic Association 7(4), 2009, pp. 858-885. federalreserve.gov
12. U.S. Bureau of Economic Analysis, NIPA Table 1.1.5, gross domestic product in current dollars; $32.49 trillion at a seasonally adjusted annual rate in the second quarter of 2026 (second estimate, August 26, 2026). bea.gov
13. Mark Doms, “The Boom and Bust in Information Technology Investment,” FRBSF Economic Review 2004, pp. 19-34; real IT investment growth of 23.8% a year 1995 to 2000, “over 3/4 percentage point” of annual GDP growth, and a 0.4 point subtraction in 2001. frbsf.org
14. Board of Governors of the Federal Reserve System, ten-year Treasury constant maturity rate, annual averages, via FRED series GS10. fred.stlouisfed.org
15. Board of Governors of the Federal Reserve System, “Open Market Operations,” historical federal funds target changes; 4.75% in June 1999 to 6.50% in May 2000. federalreserve.gov
16. Kurt G. Lunsford, “Productivity Growth and Real Interest Rates in the Long Run,” Federal Reserve Bank of Cleveland, Economic Commentary 2017-20 (November 2017). clevelandfed.org
17. James D. Hamilton, Ethan S. Harris, Jan Hatzius, and Kenneth D. West, “The Equilibrium Real Funds Rate: Past, Present, and Future,” IMF Economic Review 64(4), 2016, pp. 660-707; originally U.S. Monetary Policy Forum, 2015. econweb.ucsd.edu
18. Carlos Carvalho, Andrea Ferrero, Felipe Mazin, and Fernanda Nechio, “Underlying Trends in the U.S. Neutral Interest Rate,” FRBSF Economic Letter 2025-10 (April 21, 2025). frbsf.org
19. Isaiah Andrews and Maryam Farboodi, “Do Markets Believe in Transformative AI?” NBER Working Paper 34243 (September 2025, revised July 2026); 30 releases, four labs, “average declines in excess of ten basis points” persisting roughly 30 trading days. nber.org
20. Lukasz Rachel, “What Next for r*? A Capital Market Equilibrium Perspective on the Natural Rate of Interest,” Brookings Papers on Economic Activity (Fall 2025); the AI scenario assumes TFP growth 0.75 point higher for a decade and raises the safe natural rate “by over a percentage point.” brookings.edu
21. Daron Acemoglu, “The Simple Macroeconomics of AI,” NBER Working Paper 32487 (2024), published in Economic Policy 40(121), 2025; TFP gains of no more than 0.66% over ten years. nber.org
22. Philip R. Lane, “AI and monetary policy,” speech at the ESCB ChaMP Research Network closing conference, Rome, European Central Bank (July 6, 2026), delivered by Philipp Hartmann. ecb.europa.eu
23. Iñaki Aldasoro, Sebastian Doerr, Leonardo Gambacorta, and Daniel Rees, “The impact of artificial intelligence on output and inflation,” BIS Working Papers No. 1179 (April 2024). bis.org

Author disclosure

Educational content, not investment, tax, or lending advice. The calculator is a sensitivity tool built from published coefficients and is not a forecast of interest rates. Interest-rate outcomes are uncertain; confirm loan terms with your lender before acting.